Dran

As someone running diffusion and autoregressive models both in my basement and at work, the drive to create open-source models as a method of competition certainly makes me feel like I’m winning as a consumer.

Also, the first five digits were the same between the two numbers. Meta is guilty, but they’re guilty of grifting, not of giving a rogue AI access to some shadow database of personal details… yet? Lol

99% of the waiting time in my case is either waiting for file copies or waiting on SAP programs to run.

I wish I had low hanging fruit like that to go after.

I copy the install media locally. Although there is probably a noticable performance hot to running my main VM disk over the network.

Building a fully functional SAP system just takes that long in raw install time when your process also includes a sufficiently large system copy, and your hardware isn’t bleeding-edge.

It’s a massive application stack

I wrote and maintain a zero-to-working SAP HANA/S4 installer in pure bash.

It takes a redhat compatible from base install to a working, production-ready SAP system in about 5 hours.

It’s like ~9,000 lines of bash

An example of this:

Bitcoin mining started on cpus, then moved to gpus, and now exists on dedicated asics.

A $200 GPU vs a $200 ASIC, the ASIC is going to be a faster sha256 calculator

A $2000 GPU vs a $200 ASIC, the GPU is going to be a faster sha256 calculator

A $200 GPU from today vs a $200 ASIC from 10 years ago vs a $200 CPU from today?.. You get the idea.

There’s no way to know without specific details which will be faster. You could be running software encryption on a raspberry pi from 5 years ago or the drive could be running an encryption ASIC from 10 years ago, etc

The short answer is that: all other things being equal, it will always be faster and cheaper to do things dedicated in hardware. Comparing one implementation to another, however, is always going to be an “it depends”